Abstract: To address the inconsistency due to different factor selection in the assessment of debris flow susceptibility, and the inadequate accuracy in susceptibility evaluation due to current neural networks inefficiently extracting debris flow features, we proposed a novel approach based on a self-calibration prototype network for debris flow susceptibility assessment.Taking the valley watershed as the evaluation unit, DEM data, high-resolution imagery from Gaofen 1,and Google Earth remote sensing images were extracted as training data.The prototype network′s feature extractor was constructed with an attention mechanism and an atrous space convolution pooling pyramid structure.Additionally, a self-calibration method was employed to optimize the prototype network′s computation.Subsequently, non-debris flow valley images were inputted into the refined model, and the debris flow susceptibility index was computed to obtain the debris flow assessment level.We applied this model to evaluate valleys in the Nujiang Prefecture and compare the results with historical disaster data, showing a remarkable classification accuracy of 86.32%.The assessment outcomes of prone and high-prone areas closely aligned with the spatial distribution of historical debris flow valleys.In comparison to traditional evaluation methods, this approach demonstrates superior capability in autonomously learning debris flow features from remote sensing images and rapidly identifying and assessing disaster-prone regions.This methodology can provide a fresh perspective for advancing research on debris flow disasters by deep learning.